Pseudo-Label Refinement for Robust Wheat Head Segmentation via Two-Stage Hybrid Training

This extended abstract details our solution for the Global Wheat Full Semantic Segmentation Competition. We developed a systematic self-training framework. This framework combines a two-stage hybrid training strategy with extensive data augmentation. Our core model is SegFormer with a Mix Transformer (MiT-B4) backbone. We employ an iterative teacher-student loop. This loop progressively refines model accuracy. It also maximizes data utilization. Our method achieved competitive performance. This was evident on both the Development and Testing Phase datasets.

💡 Research Summary

The paper presents a comprehensive solution for the Global Wheat Full Semantic Segmentation Competition, built around a self‑training framework that leverages both limited ground‑truth annotations and abundant unlabeled images. The backbone of choice is SegFormer with a Mix‑Transformer (MiT‑B4) encoder, selected for its ability to capture long‑range dependencies while remaining computationally efficient. The authors keep the architecture constant across all training phases, varying only the data and hyper‑parameters to suit each stage.

The training pipeline consists of three logical stages. Stage 0 trains an initial “teacher” model on the 99 available manually annotated wheat‑head masks. This teacher generates pseudo‑labels for the large pool of unlabeled images; a confidence threshold filters out low‑quality predictions, ensuring that only high‑certainty masks are retained. Stage 1 (pseudo‑label pre‑training) uses these filtered masks to train a student model from scratch on 512 × 512 images for 40 epochs (batch = 8, LR = 6 × 10⁻⁵). This low‑resolution phase encourages the network to learn robust, global features from a large, noisy dataset. Stage 2 (fine‑tuning) then refines the same model on the true‑label set at full resolution (1024 × 1024) for 25 epochs per fold (batch = 1, LR = 1 × 10⁻⁵) using the AdamW optimizer. The fine‑tuning stage captures fine‑grained details such as wheat‑head boundaries that are essential for high IoU scores.

Data augmentation is heavily employed throughout both stages: random cropping, horizontal/vertical flips, 90° rotations, scaling (±20 %), translation (±6.25 %), additional rotations (±30 °), brightness/contrast adjustments (±30 %), HSV jitter, and ImageNet‑style normalization. These transformations increase the effective size of the training set and improve robustness to field variability (lighting, viewpoint, and plant morphology).

A key novelty is the iterative teacher‑student loop. After Stage 2, the refined student becomes the new teacher, generating an improved set of pseudo‑labels for the next iteration. This cyclical process is repeated multiple times, progressively enhancing label quality and model performance without requiring additional manual annotations. Within each iteration, the authors perform 10‑fold cross‑validation, yielding ten independently trained models. During inference, they employ model ensembling by averaging the logits of all folds, and they further apply extensive test‑time augmentation (original image, horizontal/vertical flips, 90° rotations, and multi‑scale inputs at 0.75× and 1.25×). Augmented logits are transformed back to the original orientation, averaged, up‑sampled to 1024 × 1024, and finally arg‑maxed to produce the segmentation mask, which is then resized to the required 512 × 512 output format.

The reported results demonstrate the effectiveness of this approach: a mean Intersection‑over‑Union (mIoU) of 0.7480 on the development set and 0.7099 on the hidden test set. Qualitative examples show consistent wheat‑head detection across diverse field conditions, indicating that the combination of low‑resolution pre‑training, high‑resolution fine‑tuning, and iterative pseudo‑label refinement successfully balances global feature learning with precise boundary recovery.

The authors acknowledge several limitations. The backbone is fixed to SegFormer‑MiT‑B4, so the impact of alternative transformer‑based encoders remains unexplored. The confidence threshold for pseudo‑label selection is manually set, which may not be optimal for all datasets. Future work could investigate adaptive thresholding, backbone diversification, and more sophisticated multi‑scale feature fusion to further boost performance. Nonetheless, the presented framework offers a practical blueprint for leveraging massive unlabeled agricultural imagery in a self‑supervised manner, and it sets a strong baseline for future wheat‑head segmentation research.